uv_uptime returns incorrect values on OpenVZ containers.

[schamberg97]

• Version: All current versions, to the best of my knowledge
• Platform: Any Linux container in OpenVZ/Virtuozzo, e.g.: Linux vps1607355603 4.15.0 #1 SMP Tue Aug 25 11:59:26 MSK 2020 x86_64 x86_64 x86_64 GNU/Linux
• Related Node.js issue: os.uptime returns the wrong value nodejs/node#36244 (for our exploration of the issue)

libuv/src/unix/linux-core.c

Lines 606 to 628 in 9c6cec8

	int uv_uptime(double* uptime) {
	static volatile int no_clock_boottime;
	struct timespec now;
	int r;
	/* Try CLOCK_BOOTTIME first, fall back to CLOCK_MONOTONIC if not available
	* (pre-2.6.39 kernels). CLOCK_MONOTONIC doesn't increase when the system
	* is suspended.
	*/
	if (no_clock_boottime) {
	retry: r = clock_gettime(CLOCK_MONOTONIC, &now);
	}
	else if ((r = clock_gettime(CLOCK_BOOTTIME, &now)) && errno == EINVAL) {
	no_clock_boottime = 1;
	goto retry;
	}
	if (r)
	return UV__ERR(errno);
	*uptime = now.tv_sec;
	return 0;
	}

Hey there :) It appears that libuv in OpenVZ environments returns incorrect values. For instance, if one were to check the uptime using cat /proc/uptime and Node.js's os.uptime(), one would get the following results respectively: 3187731.82 and 6740816. In other words, libuv returns results that are higher than container's uptime. Same erroneous result is achieved with this code provided by @gireeshpunathil:

#include <time.h>
#include <errno.h>
#include <stdio.h>

int main() {
  struct timespec now;
  int r;

  r = clock_gettime(CLOCK_BOOTTIME, &now);
  if (r != 0) {
    fprintf(stderr, "clock_gettime failed with %d\n", errno);
    return errno;
  }
  fprintf(stderr, "system uptime: %ld\n", now.tv_sec);
  return 0;
}

To get into the root of the problem, we need to understand how OpenVZ operates. OpenVZ consists of a host Linux system with highly patched kernel, which then runs the containers. Unfortunately, all containers share the host kernel, but are provided with a separate simulation of /proc/ (procfs) filesystem for each container. Thus, the earlier mentioned 6740816 uptime from libuv coincides with the uptime of the physical host system.

In practice, this means that the clock_gettime is inherently unreliable on OpenVZ systems. It is suggested by @davedodo on a related node.js issue nodejs/node#36244 (comment) that sysinfo function works correctly with OpenVZ systems (like in Cygwin implementation). However, there might be other alternatives that I (we?) are unaware of.

It is my understanding that this issue concerns only OpenVZ/Virtuozo. Thus, in order to check, whether the system is an OpenVZ container, it appears sufficient to check that /proc/vz/veinfo file exists at launch time (it contains container ID, public IP of current container and current number of running processes).

I am not providing any PR at this point, because this issue may involve quite a bit of decision making on this issue, with a non-exhausting list of three alternatives:

1. Rewrite uv_uptime entirely using sysinfo for all cases
2. Check for OpenVZ at launch time to avoid clock_gettime in uv_uptime on OpenVZ and use sysinfo instead
3. Ignore the issue

[bnoordhuis]

Option 1 is probably off the table.

In 3a1be72 and 1c22b44 I switched some things from sysinfo() to /proc because of compatibility issues with lxc, another virtualization solution.

I think (not 100% sure) lxc intercepts clock_gettime() by remapping the vDSO - meaning it Just Works from libuv's perspective - but it does no such thing for sysinfo().

What could probably work - and is virtualization agnostic - is to read the first field from /proc/uptime if that file exists (the second field is the system idle time) with a fallback to clock_gettime(). Pull request welcome.

Parsing a file is much more expensive than a vDSO "system call" (it's usually serviced from the vDSO without making an actual system call) but I don't think that matters for uv_uptime(), it's not performance critical.

[schamberg97]

Option 1 is probably off the table.

In 3a1be72 and 1c22b44 I switched some things from sysinfo() to /proc because of compatibility issues with lxc, another virtualization solution.

I think (not 100% sure) lxc intercepts clock_gettime() by remapping the vDSO - meaning it Just Works from libuv's perspective - but it does no such thing for sysinfo().

What could probably work - and is virtualization agnostic - is to read the first field from /proc/uptime if that file exists (the second field is the system idle time) with a fallback to clock_gettime(). Pull request welcome.

Parsing a file is much more expensive than a vDSO "system call" (it's usually serviced from the vDSO without making an actual system call) but I don't think that matters for uv_uptime(), it's not performance critical.

That’s how I saw the solution as well initially, but I had some aversion to it, as it is more expensive. I could make a PR today, if I am not disturbed by other matters. Otherwise, I would have no problem if someone else does it.

[bnoordhuis]

I'd say go for it. :-)